摘要
针对罂粟检测任务模型存在的小目标识别、背景复杂、目标物体相对尺度变化等问题,提出了一种基于YOLOv8n的罂粟识别改进算法。该算法通过引入iRMB注意力机制以增强罂粟小目标检测能力,通过将upsample模块替换为CARAFE上采样算子以提升罂粟在复杂背景中的识别率,通过将损失函数由CIoU替换为MPDIoU以应对罂粟遥感目标相对尺度变化问题。研究结果表明:基于YOLOv8n的罂粟识别改进算法可将传统YOLOv8n算法的mAP从83.1%提升至86.6%,其中罂粟果实识别的AP提升1.0个百分点,罂粟花蕊识别的AP提升6.1个百分点,实现了对罂粟识别综合性能的提升。
Aiming at the problems of small object recognition,complex background,and relative scale of target object change in poppy detection task model,a poppy recognition algorithm improvement strategy based on YOLOv8n was proposed.In improved algorithm,firstly,the iRMB attention mechanism was introduced to enhance the ability to detect small targets of poppy.Secondly,the upsample module was replaced with the CARAFE upsampling operator to improve the recognition rate of poppy in complex backgrounds.Finally,the loss function was replaced from CIoU to MPDIoU to cope with the relative scale variation of remote sensing targets of poppy.The study found that the improved poppy recognition algorithm based on YOLOv8n can increase the mAP of traditional YOLOv8n algorithm from 83.1% to 86.6%,of which the AP of poppy fruit recognition increased by 1.0 percentage point,and AP of poppy stamen recognition increased by 6.1 percentage points.The improved YOLOv8n recognition algorithm enhanced the comprehensive performance of poppy recognition.
作者
陈海涛
王辉
邓涛
刘永粤
张琪
CHEN Haitao;WANG Hui;DENG Tao;LIU Yongyue;ZHANG Qi(Graduate School,China People s Police University,Langfang Hebei 065000,China;Hebei Key Laboratory of Low-Altitude Safety Management and Control Technology,Langfang Hebei 065000,China;SWAT Detachment,Changzhou Public Security Bureau,Changzhou Jiangsu 213022,China;School of Physical Science and Technology,Southwest University,Chongqing 400715,China;Chongqing Key Laboratory of Micro&Nano Structured Optoelectronics,Southwest University,Chongqing 400715,China)
出处
《西南大学学报(自然科学版)》
北大核心
2025年第6期201-212,共12页
Journal of Southwest University(Natural Science Edition)
基金
国家重点研发计划项目(2023YFB2905403)
河北省重点研发计划项目(20375601D)
重庆市自然科学基金面上项目(CSTB2023NSCQ-MSX0120)。
